"Every Agentic Engineering Hack I Know (June 2026)" — Matt Van Horn
Sequel to his March "Every Claude Code Hack I Know" (913K views). 22-section practitioner stack. Founder shared 2026-06-03 with the read: we already do most of this. This note is the hack-by-hack parity map confirming that, plus the genuine gaps and the deliberate divergences.
Why this is in the vault
It's a credible-practitioner audit of RDCO's own agentic-engineering stack. Van Horn ships real agent tooling (last30days, Printing Press, Agent Cookie), so running his 22-hack list against the COO-agent setup is a cheap external check on whether RDCO is current. The headline finding is reassuring (parity or ahead on most hacks), but the value is the two genuine deltas it surfaces: hack #7 (an email address for Claude Code) as a third pattern for the open phData work-agent email path, and hack #11 (raw transcript into the agent) as the durable insight behind the founder's Granola interest. Filed so those two action candidates and the founder's "we already do most of this" verdict are on record.
Mapping against Ray Data Co
The body is itself the mapping: a hack-by-hack parity table scoring each Van Horn practice against RDCO's equivalent, followed by the two deltas worth acting on. Read it as a coverage audit, not new doctrine.
Parity map: Van Horn hack → RDCO equivalent
| # | Hack | RDCO status |
|---|---|---|
| 1–3 | /ce-plan → /ce-work, plan.md as checkpoint, "plan for the plan" for deep non-code work |
Parity (different harness). Same spec→plan→build loop via build-project, the pipeline-* seats, and the implementation-notes pattern. We use our own scaffolding, not the Compound Engineering plugin. |
| 4 | Get voice-pilled (Monologue/Wispr Flow) | Gap / unknown. Founder comms are text (iMessage/Discord). Voice-to-LLM not in the stack. Low priority. |
| 5 | 4–6 cmux tabs, one task each | Partial. Channels-agent runs in tmux; parallelism is via sub-agent dispatch + cron, not manual tab juggling. |
| 6 | Terminal opens straight into Claude | Parity. Channels-agent + restart script launch directly into a named CC session. |
| 7 | Remote control + give Claude an email address (AgentMail → fresh session spawns on inbound email) | GENUINE NEW ITEM. Directly relevant to the open phData work-agent email-path decision. See below. |
| 8 | YOLO bypassPermissions / skip-permission-prompt |
Deliberate divergence. We run auto-mode but keep HARD gates on deploy/production-write/external-send (classifier hard-gate, no-autonomous-external-email). Not a gap — a safety choice. |
| 9 | Codex as a second engine (Claude plans, Codex builds) | Deliberate divergence. RDCO builds deep on the Anthropic stack (L5 north-star direction). Single-engine on purpose. |
| 10 | last30days research-before-plan |
Parity. /deep-research, /curiosity, qmd vault search. |
| 11 | Granola, drop the RAW transcript into the agent (don't summarize first) | Gap — founder's pick. No meeting capture today. The valuable part isn't Granola, it's the raw-transcript-into-agent move. See below. |
| 12 | Be the human signal (you supply taste, agents supply volume) | Strong parity — it's our operating model. = [[feedback_advisor_not_pair_programmer]], founder-is-signal. |
| 13 | HyperFrames video (video as HTML an agent writes) | Ahead. Full HyperFrames skill suite (hyperframes, -cli, -media, -registry, adapters) already in place. |
| 14 | Notes as the agent's knowledge base (Bear/Obsidian + gbrain/supermemory) | Ahead. The vault + qmd semantic search + graph.duckdb typed knowledge graph is a deeper version of exactly this. |
| 15 | Mac mini + mosh/tmux remote, autonomous remote work | Parity. Channels-agent is a Mac Mini + tmux always-on setup. |
| 16 | Proof for sharing plans with non-terminal humans | Functional parity. HQ (hq.raydata.co) generative-UI is the founder-facing surface; iMessage/Discord close the loop. |
| 17 | Write your own skills (copy a working skill's shape) | Strong parity. = [[feedback_skills_over_commands]] + /skillify. |
| 18 | Open-source contribution as a loop | N/A to current RDCO model. |
| 19 | M5 Max laptop / power bricks | N/A. |
| 20 | Printing Press: agent-native CLIs for real-life errands + Agent Cookie (browser-session auth) | Partial / different shape. We hand errands to agents via MCP servers (Stripe, Calendar, Gmail, Monarch, etc.), not generated CLIs. Agent Cookie's "real browser session, no password paste" is interesting for services without an MCP. |
| 21 | The honest part: AI psychosis / touch grass / build what people want | Human note. Worth keeping in view. |
| 22 | Article written via talk→plan→build | Meta. |
The two things worth acting on
Hack #7 — email address for Claude Code → third option for the phData email path
The open phData work-agent decision (~/rdco-vault/08-tooling/2026-05-30-phdata-work-agent-setup-plan.md) was framed as two options for the email path: claude mcp add Gmail vs. a local Gmail-API draft script. Van Horn's agentmail-to-claude-code (open source) is a third pattern: an inbox a fresh CC session spawns from on inbound email (allowlisted, DKIM/SPF-gated).
- Fit: "email the bug from your phone, a session is running before you're back at a screen" matches the MVP = email-draft-triage goal, and the allowlist + DKIM/SPF gate is a clean security posture.
- Caveat (honest): it runs as an inbox-watching daemon over WebSocket — which cuts against the founder's stated "session-based, NOT a daemon" preference for the air-gapped work machine. So it's a reference point, not a drop-in. Surface it, don't push it.
Hack #11 — the raw-transcript move (the actual value behind "Granola")
Founder is considering Granola for work recording (appeal: nothing joins the call). The durable insight is workflow, not tool: drop the raw, un-summarized transcript into the agent and let it extract against your codebase/context/prior plans. "Granola context + codebase + prior plans = gold." Any local recorder that yields a raw transcript feeds this; Granola just happens to capture without a bot participant.
Related
- [[2026-06-02-thariq-dynamic-workflows-harness-for-every-task]] (same-day share; the Anthropic-native version of the orchestration Van Horn does by hand)
- [[feedback_advisor_not_pair_programmer]] (hack #12 = our operating model)
- [[feedback_skills_over_commands]] (hack #17)
- [[project_phdata_cert_escalator_path]] / phData work-agent plan (hack #7 cross-ref)